Rapid Diagnosis of Acute Heart Disease by Cloud-based High Performance Computing for Computer Vision Oleksii Morozov Physi...
The Heart <ul><li>Life-sustaining pump: 2’500’000 L/year of vital blood </li></ul><ul><li>Coronary artery disease (CAD) is...
Cardiology yesterday <ul><li>Tools with relatively low information content </li></ul>
Cardiology today <ul><li>More tools, more information </li></ul><ul><li>Subjective decision mostly relying on experience o...
Cardiology tomorrow <ul><li>More advanced technologies </li></ul><ul><ul><li>Multidimensional information </li></ul></ul><...
Cardiac Ultrasound <ul><li>One of the modern tools for evaluation of the heart function </li></ul><ul><ul><li>HF sound wav...
Cardiac Ultrasound Diagnostic value <ul><li>Heart wall assessment </li></ul>
Cardiac Ultrasound Diagnostic value <ul><li>Pumping function </li></ul>
Cardiac Ultrasound Diagnostic value <ul><li>Valve function </li></ul>
3D Cardiac Ultrasound <ul><li>Explore heart in 3D </li></ul><ul><ul><li>Freehand ultrasound (Manual sweeping) </li></ul></ul>
3D Cardiac Ultrasound <ul><li>Explore heart in 3D </li></ul><ul><ul><li>Freehand ultrasound (Manual sweeping) </li></ul></...
3D Cardiac Ultrasound <ul><li>Explore heart in 3D </li></ul><ul><ul><li>Freehand ultrasound (Manual sweeping) </li></ul></...
Cardiac Ultrasound <ul><ul><li>Ultrasound machine = a  transducer  +  a  supercomputer </li></ul></ul>50’000 – 500’000 USD...
Computational problems  in Cardiac Ultrasound <ul><li>Signal reconstruction </li></ul>Non-uniformly sampled  measurements ...
3D+time signal reconstruction <ul><li>Inherent non-uniformity of scanning </li></ul><ul><ul><li>Spatial non-uniformity </l...
3D+time signal reconstruction <ul><li>Inherent non-uniformity of scanning </li></ul><ul><ul><li>Spatial non-uniformity </l...
3D+time signal reconstruction <ul><li>Inherent non-uniformity of scanning </li></ul><ul><ul><li>Spatial non-uniformity </l...
3D+time signal reconstruction A spline solution <ul><li>B-spline non-uniform interpolation by Arigovindan, Unser (EPFL, Sw...
3D+time signal reconstruction A spline solution <ul><li>Obstacles in 3D/4D </li></ul><ul><ul><ul><li>Complexity is exponen...
3D+time signal reconstruction A spline solution <ul><li>Tensor based approach applied to ultrasound data from continuously...
Computational problems  in Cardiac Ultrasound <ul><li>Tissue/blood motion estimation </li></ul><ul><ul><li>Doppler Ultraso...
Computational problems  in Cardiac Ultrasound <ul><li>B-spline based tissue motion reconstruction </li></ul><ul><ul><li>Co...
Computational problems  in Cardiac Ultrasound <ul><li>Blood flow reconstruction </li></ul><ul><ul><li>Resolves ambiguity o...
Pathway to distributed supercomputing <ul><li>Multicore (IBM Power7) claimed 260 GFLOP/chip </li></ul><ul><li>Cluster (Uni...
Cloud Ultrasound Processing Service <ul><li>Reasons </li></ul><ul><ul><li>Processing of large multidimensional multimodal ...
Cloud Ultrasound Processing Service Cloud 4D acquisition with real-time on board visualization Interactive web-based visua...
Cloud Ultrasound Processing Service <ul><li>Record data </li></ul><ul><ul><li>Raw data </li></ul></ul><ul><ul><ul><li>180 ...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Problem is very large for solving using direct solvers -...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Iteration can be distributed relative to the grid </li><...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Data dependency </li></ul>OP{k,m}()  uses data outside t...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Data dependency </li></ul><ul><ul><li>Data size 512 x 51...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Computational load </li></ul><ul><ul><li>Data size 512 x...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Multiresolution </li></ul><ul><ul><li>Coarse to scale pr...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Multiresolution in solving algorithm </li></ul><ul><ul><...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Multiresolution in solving algorithms </li></ul><ul><ul>...
Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Total requirements per compute unit </li></ul><ul><ul><l...
Cloud Ultrasound Processing Service <ul><li>Costs estimation </li></ul><ul><li>Switzerland (1/1000 of world population) </...
Collaborative Research enabled by the Cloud <ul><li>Globally available service for processing, storing and accessing medic...
Conclusion <ul><li>An approach to rapid diagnosis of heart disease using cloud based distributed computing </li></ul><ul><...
Upcoming SlideShare
Loading in...5
×

Rapid Diagnosis of Acute Heart Disease by Cloud-based High ...

323

Published on

0 Comments
1 Like
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total Views
323
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
4
Comments
0
Likes
1
Embeds 0
No embeds

No notes for slide
  • The heart is a life-sustaining pump which supplies the blood to all the organs of the human. Its highly optimized structure (myocardium) and functioning allows to perform an enormous activity. At an average rate of 72 beats per minute the heart pumps about 2.5 millions liter of the vital blood per year. As the other organs supplied with blood, the heart needs its own oxygen and nutrition supply. This task is performed by the coronary artery system of the heart. Failures of this system, known as CAD, are the most frequent causes of heart malfunction and death. It is essentially the leading cause of death worldwide.
  • - Prevention - Diagnose - Treatment
  • That is why it is so important to be able to assess the functioning of the heart to foresee the problems (diagnosis) and to treat already affected hearts. Modern technology provides a number of tools for this. One of them is ultrasound imaging which is based of transmission of high frequency sound waves into the body and receiving the response from the tissue imaged objects. This tool has evolved as a well-established non-invasive technique for the evaluation of the local and global heart function and for the evaluation of the blood flow.
  • Splines offer an attractive framework for solving the problem of multidimensional non-uniform interpolation.
  • Splines offer an attractive framework for solving the problem of multidimensional non-uniform interpolation.
  • In few seconds the cloud should be ready to provide the required information!
  • In few seconds the cloud should be ready to provide the required information!
  • Rapid Diagnosis of Acute Heart Disease by Cloud-based High ...

    1. 1. Rapid Diagnosis of Acute Heart Disease by Cloud-based High Performance Computing for Computer Vision Oleksii Morozov Physics in Medicine Research Group University Hospital of Basel Switzerland April 8, 2010
    2. 2. The Heart <ul><li>Life-sustaining pump: 2’500’000 L/year of vital blood </li></ul><ul><li>Coronary artery disease (CAD) is most frequent cause of heart malfunction and death </li></ul><ul><ul><li>World largest killer (WHO) </li></ul></ul><ul><ul><ul><li>~29% of global death </li></ul></ul></ul><ul><ul><ul><li>17’100’000 lives/year </li></ul></ul></ul>
    3. 3. Cardiology yesterday <ul><li>Tools with relatively low information content </li></ul>
    4. 4. Cardiology today <ul><li>More tools, more information </li></ul><ul><li>Subjective decision mostly relying on experience of a doctor </li></ul>
    5. 5. Cardiology tomorrow <ul><li>More advanced technologies </li></ul><ul><ul><li>Multidimensional information </li></ul></ul><ul><ul><li>High quality, high resolution data </li></ul></ul><ul><ul><li>Multimodal information </li></ul></ul><ul><ul><li>Quantitative, objective, integrative computer based analysis </li></ul></ul><ul><ul><li>Worldwide-networked standards and databases </li></ul></ul><ul><li>Problems </li></ul><ul><ul><li>Need for high performance computing in a distributed environment but only for a fraction of the time </li></ul></ul><ul><ul><li>Global storage network for storing large datasets </li></ul></ul>
    6. 6. Cardiac Ultrasound <ul><li>One of the modern tools for evaluation of the heart function </li></ul><ul><ul><li>HF sound waves : No Radiation/Ionization </li></ul></ul><ul><ul><li>Safe, Non-invasive, Fast, Portable, Cheap </li></ul></ul><ul><ul><li>“ -” Rather low signal to noise ratio </li></ul></ul>
    7. 7. Cardiac Ultrasound Diagnostic value <ul><li>Heart wall assessment </li></ul>
    8. 8. Cardiac Ultrasound Diagnostic value <ul><li>Pumping function </li></ul>
    9. 9. Cardiac Ultrasound Diagnostic value <ul><li>Valve function </li></ul>
    10. 10. 3D Cardiac Ultrasound <ul><li>Explore heart in 3D </li></ul><ul><ul><li>Freehand ultrasound (Manual sweeping) </li></ul></ul>
    11. 11. 3D Cardiac Ultrasound <ul><li>Explore heart in 3D </li></ul><ul><ul><li>Freehand ultrasound (Manual sweeping) </li></ul></ul><ul><ul><li>Mechanical sweeping ultrasound (Motor driven) </li></ul></ul>
    12. 12. 3D Cardiac Ultrasound <ul><li>Explore heart in 3D </li></ul><ul><ul><li>Freehand ultrasound (Manual sweeping) </li></ul></ul><ul><ul><li>Mechanical sweeping ultrasound (Motor driven) </li></ul></ul><ul><ul><li>Live 3D ultrasound (2D arrays with electrical sweeping) </li></ul></ul>
    13. 13. Cardiac Ultrasound <ul><ul><li>Ultrasound machine = a transducer + a supercomputer </li></ul></ul>50’000 – 500’000 USD <ul><ul><li>Idle 90% of the time </li></ul></ul>
    14. 14. Computational problems in Cardiac Ultrasound <ul><li>Signal reconstruction </li></ul>Non-uniformly sampled measurements Complete gridded or continuous data representation
    15. 15. 3D+time signal reconstruction <ul><li>Inherent non-uniformity of scanning </li></ul><ul><ul><li>Spatial non-uniformity </li></ul></ul><ul><ul><li>Serialism in scanning </li></ul></ul>
    16. 16. 3D+time signal reconstruction <ul><li>Inherent non-uniformity of scanning </li></ul><ul><ul><li>Spatial non-uniformity </li></ul></ul><ul><ul><li>Serialism in scanning </li></ul></ul><ul><li>Non-uniformity in synchronization (ECG) </li></ul>
    17. 17. 3D+time signal reconstruction <ul><li>Inherent non-uniformity of scanning </li></ul><ul><ul><li>Spatial non-uniformity </li></ul></ul><ul><ul><li>Serialism in scanning </li></ul></ul><ul><li>Non-uniformity in synchronization (ECG) </li></ul><ul><li>Body motion artifacts (breathing) </li></ul>4D non-uniform data I(x,y,z,t) ?
    18. 18. 3D+time signal reconstruction A spline solution <ul><li>B-spline non-uniform interpolation by Arigovindan, Unser (EPFL, Switzerland 2005) </li></ul><ul><ul><li>Robust global interpolation: handles oversampling and undersampling (gaps) in the data </li></ul></ul><ul><ul><li>Sparse and well-conditioned alternative to the optimal RBF solution </li></ul></ul><ul><ul><li>Enjoys multiresolution properties (way to fast solving) </li></ul></ul><ul><ul><li>Parallelizability of solving process </li></ul></ul><ul><ul><li>Successfully applied to 2D problems </li></ul></ul>
    19. 19. 3D+time signal reconstruction A spline solution <ul><li>Obstacles in 3D/4D </li></ul><ul><ul><ul><li>Complexity is exponentially dependent on the data size 128 x 128 x 128 x 18 –> 78’752’009’856 non-zeros (312 Gbyte in single precision) </li></ul></ul></ul><ul><li>Tensor based approach by Morozov, Hunziker, Unser 2009 </li></ul><ul><ul><li>Tensor decomposition of the problem </li></ul></ul><ul><ul><li>Relaxed storage requirements </li></ul></ul><ul><ul><li>Feasibility on standard workstations </li></ul></ul><ul><ul><ul><li>~9 millions of measurements with size 128 x 128 x 128 x 18 -> 30 minutes on my dual core laptop </li></ul></ul></ul>
    20. 20. 3D+time signal reconstruction A spline solution <ul><li>Tensor based approach applied to ultrasound data from continuously rotating transducer </li></ul>
    21. 21. Computational problems in Cardiac Ultrasound <ul><li>Tissue/blood motion estimation </li></ul><ul><ul><li>Doppler Ultrasound imaging (State of the art) </li></ul></ul><ul><ul><ul><li>Semi-quantitative measurements </li></ul></ul></ul><ul><ul><li>Full motion reconstruction </li></ul></ul><ul><ul><ul><li>Generalization of B-spline reconstruction to vector valued data (Arigovindan, Unser 2005) </li></ul></ul></ul><ul><ul><ul><li>Employing additional constraints from physics of fluids (incompressibility, Navier-Stokes equations) </li></ul></ul></ul>
    22. 22. Computational problems in Cardiac Ultrasound <ul><li>B-spline based tissue motion reconstruction </li></ul><ul><ul><li>Continuous </li></ul></ul><ul><ul><li>Fully quantifiable </li></ul></ul><ul><ul><li>Can be combined with Doppler for better robustness </li></ul></ul>
    23. 23. Computational problems in Cardiac Ultrasound <ul><li>Blood flow reconstruction </li></ul><ul><ul><li>Resolves ambiguity of Doppler measurements </li></ul></ul><ul><ul><li>Continuous </li></ul></ul><ul><ul><li>Fully quantifiable </li></ul></ul>
    24. 24. Pathway to distributed supercomputing <ul><li>Multicore (IBM Power7) claimed 260 GFLOP/chip </li></ul><ul><li>Cluster (UniBasel) 34’500 GFLOP/400 cores </li></ul><ul><li>GPGPU (ATI 4870X2) 2’000 GFLOP/card </li></ul><ul><li>GPGPU array </li></ul><ul><li>FPGA accelerator cards: dozens of GFLOP/chip, up to 512 chips per system, low power </li></ul><ul><ul><li>In exploration within ICES Microsoft project </li></ul></ul><ul><li>Cloud - Microsoft Azure </li></ul>
    25. 25. Cloud Ultrasound Processing Service <ul><li>Reasons </li></ul><ul><ul><li>Processing of large multidimensional multimodal medical data requires vast computational power </li></ul></ul><ul><ul><li>Building/maintaining own HPC infrastructure is overly expensive </li></ul></ul><ul><ul><li>Relatively rare use of HPC power (few times per day) </li></ul></ul><ul><ul><li>Availability at multiple points of care (medical practices and hospital emergency rooms) </li></ul></ul><ul><ul><li>Unified storage/access of the multimodal medical data </li></ul></ul>
    26. 26. Cloud Ultrasound Processing Service Cloud 4D acquisition with real-time on board visualization Interactive web-based visualization of the result Rendered images and quantitative information Visualization/Analysis parameters Record data User
    27. 27. Cloud Ultrasound Processing Service <ul><li>Record data </li></ul><ul><ul><li>Raw data </li></ul></ul><ul><ul><ul><li>180 beams x 500 samples x 100 frames x 10 sec -> 85 Mb </li></ul></ul></ul><ul><ul><ul><li>Additional information (geometry) -> few Kb </li></ul></ul></ul><ul><ul><li>Lossless compressed DICOM </li></ul></ul><ul><ul><li>Low latency response to the user by sending first a subpart of the data for coarser resolution reconstruction </li></ul></ul>
    28. 28. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Problem is very large for solving using direct solvers -> use iterative solver </li></ul>C i+1 = C i + OP(C i ) OP – linear operator 2 iterations 50 iterations 80 iterations
    29. 29. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Iteration can be distributed relative to the grid </li></ul>C{1,1} i+1 = C{1,1} i + OP{1,1}(C i ) C{1,2} i+1 = C{1,2} i + OP{1,2}(C i ) C{2,1} i+1 = C{2,1} i + OP{2,1}(C i ) C{2,2} i+1 = C{2,2} i + OP{2,2}(C i ) C{k,m} – solution subpart dedicated to a compute unit OP{k,m}() – operator applied by a {k,m}’s compute unit dx dy dx, dy – grid spacing Completely independent output C{1,1} C{1,2} C{2,1} C{2,2}
    30. 30. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Data dependency </li></ul>OP{k,m}() uses data outside the bounds of C{k,m} C{k,m} <ul><ul><li>Extents of dependent input data: 3 samples for cubic spline </li></ul></ul><ul><ul><li>At each iteration this data is transferred among adjacent units </li></ul></ul><ul><ul><li>Performance limiting factor </li></ul></ul>
    31. 31. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Data dependency </li></ul><ul><ul><li>Data size 512 x 512 x 512 x 64 </li></ul></ul><ul><ul><li>Single precision: 32 GB </li></ul></ul><ul><ul><li>Infiniband QDR 12X ( ~12GB/s ) </li></ul></ul>Number of units Size of dependent data per unit, MB Total data transfers for single iteration, MB Maximal number of iterations/s (excluding CPU time) 64 72 4608 166 128 48 6144 250 256 30 7680 400 512 18 9216 666 1024 12 12288 1000
    32. 32. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Computational load </li></ul><ul><ul><li>Data size 512 x 512 x 512 x 64 </li></ul></ul><ul><ul><li>Intel® Quad Core 2.67 GHz (~30 GFLOP/s in single precision) </li></ul></ul><ul><ul><li>PC3-10600 DDR3-SDRAM (30 GB/s) </li></ul></ul><ul><ul><li>115’000’000 data samples </li></ul></ul><ul><ul><li>Total requirements: ~5000 GFLOP, ~3000 GB of memory transfers </li></ul></ul>Number of units Maximal number of iterations/s (including inter-unit communication) 64 0.24 128 0.48 256 0.96 512 1.92 1024 3.84
    33. 33. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Multiresolution </li></ul><ul><ul><li>Coarse to scale propagation – getting general from coarser scales and improving details on finer scales </li></ul></ul><ul><ul><li>Inherent spline inter-scale relation </li></ul></ul>
    34. 34. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Multiresolution in solving algorithm </li></ul><ul><ul><li>Coarse to scale propagation – getting general from coarser scales and improving details on finer scales </li></ul></ul><ul><ul><li>Inherent spline inter-scale relation </li></ul></ul><ul><ul><li>Multigrid solving algorithm </li></ul></ul>Iterate Iterate Iterate Direct solve Iterate Scale 1 Scale 2 Scale N Projection to coarser scale Projection to finer scale
    35. 35. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Multiresolution in solving algorithms </li></ul><ul><ul><li>Coarse to scale propagation – getting general from coarser scales and improving details on finer scales </li></ul></ul><ul><ul><li>Inherent spline inter-scale relation </li></ul></ul><ul><ul><li>Multigrid solving algorithm </li></ul></ul><ul><ul><ul><li>Few iterations needed at each scale to get reasonably good solution </li></ul></ul></ul><ul><ul><ul><li>With each coarser scale the cost of iteration decreases exponentially </li></ul></ul></ul><ul><ul><ul><li>In total requires much less computational load than pure iteration </li></ul></ul></ul>
    36. 36. Cloud Ultrasound Processing Service Signal reconstruction <ul><li>Total requirements per compute unit </li></ul><ul><ul><li>Data size 512 x 512 x 512 x 64 </li></ul></ul><ul><ul><li>6 scales including finest scale </li></ul></ul><ul><ul><li>16 iterations at each scale </li></ul></ul><ul><li>Motion reconstruction algorithms require ~9 times more computational load </li></ul>Number of units Memory, GB CPU time, s 64 2.7 62 128 1.36 31.2 256 0.68 15.6 512 0.34 7.8 1024 0.17 3.9
    37. 37. Cloud Ultrasound Processing Service <ul><li>Costs estimation </li></ul><ul><li>Switzerland (1/1000 of world population) </li></ul><ul><ul><li>700 cardiologists/7’000’000 population </li></ul></ul><ul><ul><li>1000 echocardiograms per year per cardiologist </li></ul></ul><ul><ul><li>Multiple views(3) per patient </li></ul></ul><ul><ul><li>Multiple analyses (3) per view </li></ul></ul>7’000’000 use cases/year 3043 use cases/hour Full loaded 3 x 1024 instances (Uniform load in Switzerland) Data size 512x512x512x64 Storage, GB 32 Number of instances 1024 Compute hours 0.0011 hour∙instance 1.13
    38. 38. Collaborative Research enabled by the Cloud <ul><li>Globally available service for processing, storing and accessing medical data </li></ul><ul><li>Standardized DICOM interface for unifying data access </li></ul><ul><li>Involve all interested parties around the world </li></ul><ul><li>World-wide large scale trials are possible </li></ul><ul><li>Getting more statistics for rare cases </li></ul><ul><li>Building reference datasets for known cases </li></ul>
    39. 39. Conclusion <ul><li>An approach to rapid diagnosis of heart disease using cloud based distributed computing </li></ul><ul><ul><li>Replace ultrasound machine’s supercomputer by a cloud service for remote processing and storage </li></ul></ul><ul><ul><li>Miniaturization of the medical equipment and decrease of its costs </li></ul></ul><ul><ul><li>Availability of advanced analysis technologies for objective analysis </li></ul></ul><ul><ul><li>Availability at multiple points of care </li></ul></ul><ul><ul><li>Unified storage and access of medical data </li></ul></ul><ul><ul><li>Enables collaborative research </li></ul></ul>
    1. A particular slide catching your eye?

      Clipping is a handy way to collect important slides you want to go back to later.

    ×